“Svetlana” wake word model for ESP32 and ESPHome
microWakeWord model for the wake word “Svetlana”, trained on microwakeword.com from synthetic speech (Czech, Italian, Polish). It runs on-device on ESP32-S3 boards with ESPHome's micro_wake_word component, including the Home Assistant Voice Preview Edition, with no cloud involved.
Model facts
- Wake word
- “Svetlana”
- Language of the training voices
- Czech, Italian, Polish
- Trained
- 2026-06-12 · microWakeWord v2
- Training steps
- 20,000
- Recall, with background noise
- 54.6%
- False activations on the benchmark set
- 1.89 per hour at probability cutoff 0.75
- Sliding window
- 5
- File
- quantized INT8 TFLite, 59 KB, with a micro_wake_word manifest
- Downloads
- 3
Add it to your ESPHome device
ESPHome fetches the model straight from this site when it compiles the firmware (ESPHome 2024.7 or newer); nothing to download.
micro_wake_word:
models:
- model: https://microwakeword.com/api/models/129/manifest.jsonSign in on this page to download the ZIP (TFLite + manifest) for offline use. Models are licensed for personal and non-commercial use by default (commercial license).
Similar wake words in the library
- “Hej houmí” — 84.6% recall
- “saro hledej” — 80.4% recall
- “saro doma” — 89.2% recall
- “svetlana” — 15.0% recall
- “Zuzana” — 74.0% recall
- “sáro” — 69.8% recall
Custom wake words are trained in the Training Center; the library lists every public model with the same measurements.